Indicators for Situation Awareness

1.2 Theoretical Framework 11 people it met in order to increase the SA of the robot (Mykoniatis, Angelopoulou, Schaefer,

& Hancock, 2013). While the algorithms performed well, it is unclear from the study how the robot utilizes this information and what impact its SA has on interactions.

Even though technologies enable robots (and other technologies) to have access to an increasing number of sensors, there is only a relatively small body of research on how a robot can display its awareness to its human communication partners, and how these displays affect communication partners’ perceptions and behaviors.

features are not only limited to talk, but can come in the form of multiple modalities (Hazel, Mortensen, & Rasmussen, 2014; Lindwall & Ekström, 2012; Mondada, 2009a;

2009b). Any one action never stands alone, but relies on context in both production and evaluation. This embeddedness between context (again, in an ethnomethodological sense) and individual actions is contingency. Interaction is thus “characterized by contingency at virtually every point” (Schegloff, 1996).

This conceptualization of contingency in CA differs somewhat from what is considered

‘contingency’ in HRI. Here, contingency is understood as a linear temporal relationship, in which behavior is influenced by a stimulus (Chu, Bullard, & Thomaz, 2014; Gold &

Scassellati, 2006; Lohan et al., 2011). In HRI, contingency is understood in terms of cause and effect. This understanding stems from research in feedback models for infants (Gergely & Watson, 1999) in which researchers investigate the temporal relationship between stimulus and response in human infants. Therefore, the problem of contingency in HRI is often one of detection. The problem is solved by implementing sensors in robots that endow them with the ability to detect changes in human behavior to which they can produce a response. Thus, contingency in HRI is a feature that can be implemented, while in CA it is an ever-present feature of interaction. This is also part of the reason why analyses in CA are always presented on a case-by-case basis; the contingencies at play in any given interaction, and the understandings of context displayed, are individual. Only through careful detailed analysis can analysts uncover exactly what aspects of context communication partners attend to.

Both models of contingency rely on indicators for situation awareness. In both models, communication partners display an awareness to each others’ action in their own conduct.

The model of contingency, as understood in CA, explicates, on a very detailed level, the intricacies of social interaction, by making clear what aspects (context) of the interaction people attend to and how it affects their own conduct. As such, this model of contingency shows how interaction is accomplished and what resources people put to use in this accomplishment. The model used in HRI does not exhibit the same complexity, but therein perhaps lies its strength. Contingency, according to this model, is relatively easy to implement and operationalize, and the model also makes clear what features of conduct communication partners display an awareness towards.

Incrementality

Verbal communication is, as Schlangen and Skantze (2011) point out, almost always incremental. Participants in conversation produce speech in real-time, often without having a complete plan of what they are going to say or do during their turn (Brennan, 2000;

Levinson, 2016;Skantze & Hjalmarsson, 2010). In interaction, people do not have fully formalized plans of actions before they carry out those actions. Rather, people constantly adapt to their surroundings and produce actions on-the-fly or change provisional plans as they see the need. For example,Suchman (1987)showed that people produce interactional contributions in a piecemeal fashion. That is, they produce their contributions (speech, gesture, etc.) in small chunks, constantly updating their contributions based on what is needed to accomplish the interaction. Evidence of this behavior is found, for example, in

1.2 Theoretical Framework 13 word searches and floor-keeping devices. These features would simply not be so prevalent in social interaction if people did not produce and process action incrementally. From a CA perspective, incrementality is uncontroversial (Goodwin, 1979). However, only over the last decade is incrementality being discussed in the field of HRI. Thus, to date many robotic systems do not process speech incrementally, although considerable work is being done to change this (see, for exampleBaumann, Kennington, Hough, and Schlangen (2017), Schlangen and Skantze (2011), Skantze and Hjalmarsson (2010)).

Incremental processing relates both to the production and comprehension of actions in social interaction (Schlangen & Skantze, 2011). Since incrementality can be seen as evidence that communication partners design their action based on ongoing changes in the context, incrementality works as an indicator for situation awareness, in which communication partners signal to each other that they attend to each others’ behavior.

Proactivity

Proactivity relies on information what a communication partner might do next and can therefore be taken as a signal for SA. Proactivity is directed related to projectablity, which is part of Endsley’s situation awareness model (Endsley, 1988). Projectability is the third and last level of the model and uses information from the first two levels, perception and comprehension, in order to forecast what will happen next. Therefore, displaying to communication partners an awareness of what is about to happen, displays an awareness of what has happened already, what action is currently ongoing, etc. Projectability is also a resource people use in interaction as it unfolds:

Sentential constructions are capable of being analysed in the course of their production by a party/hearer able to use such analyses to project their possible directions and completion and loci. In the course of its construction, any sentential unit will rapidly (in conversation) reveal projectable directions and conclusions[...]. (Sacks et al., 1974)

However, in CA, projectability (Sacks et al., 1974) and predictability (Liddicoat, 2004) primarily concern the turn-taking mechanism. In other words, people design their utterances to make the turn project when a turn-transition is coming up. These signals can be produced lexically, prosodically, or by using gaze, gestures or any other kind of modality available to communication partners.

Projectability is also implicitly part of Clark’s (1996) model of common ground, in that when people signal to each other what their common ground is, they also signal what to expect. However, Clark never discusses projectability or proactivity specifically, other than saying that some events areanticipated products, based on people’s intention (Clark, 1996, p. 22).

In HRI, proactivity is related to intention and intention recognition. The general idea is that a robot should attempt to find out what its human communication partner is doing, and on this background produce behaviors that support the human in his or her task (Ali, Alili, Warnier, & Alami, 2009). Proactive or anticipatory behaviors are therefore indicators for a

robot’s situation awareness and are direct clues through which communication partners can infer which aspects of the interaction robots take into consideration.